B-83R2-markdown-ui-playground-mermaid-e2e-ts-fails-roughly-1-in-4
State: promoted/task
The spec was added by #1722 and asserts mermaid’s math renders through KaTeX 0.18 after the dedupe alias. It failed once at load ~20 on a post-merge full run, then passed three consecutive full runs at load 9-13, and passes alone in 6.2 s against a cold dev server with the Vite cache cleared — so the 60 s wait is not the binding constraint.
Not a merge regression: the spec exercises /viewer -> MdViewer -> MdMermaidBlock and never
touches obsidian-editor.
Strongest lead, and where whoever picks this up should start: the failing run’s Playwright trace holds only config entries and no failing action, which points at the run failing BEFORE the test body executed — dev-server startup or fixture-setup contention with six parallel workers — rather than an assertion timing out. The failing run took 1.2 m against a normal 42 s.
The timeout was deliberately NOT inflated, because the evidence says the timeout is not what binds
and raising it would hide the signal. The playground e2e moon task is runInCI: false, so this
cannot redden CI.